PCA Based Model Modification

Dependencies

• python == 3.7
• pytorch == 1.12.1
• numpy == 1.21.5
• thop == 0.1.1

Implementation

Layer-wise Quantization Process

• Full process described in quantization_conversion.ipynb
• Calculate PCA explained variance ratio from compressed output for each activation layer from section 1 and 2
• Layer-wise quantization execution depicted in section 3
  • Note that the input is quantized back and forth for corresponding quantized/non-quantized layers
• Sample Metrics can be found in section 4

Training Script

• Available for Cifar10/Cifar100 training for vgg11/13/16/19
• Details in train.py

        python train.py --pretrained --dataset cifar10 --model vgg13

Experiment

• Experiment code using layer-wise quantization process
• Details in collect_statistics.py

        python collect_statistics.py

Visualization

• graph/table extraction for quantization result is in visualization.ipynb

Demo Results

Layer-wise Quantization Result

	vgg16	Vgg19
cifar 10
cifar 100